POINT CLOUD WORKSHOP

AT SVCU 🌏.

🎮 A Comprehensive Introduction to 3D Data Analysis

👤 Presenter: Natapong Intarasuk

🎓 Institution: Chulalongkorn University

📧 Contact: int.natapong@gmail.com

🐙 GitHub: github.com/imtorrr

📋 TODAY'S AGENDA

✨

Fundamentals

Understand what point clouds are and their unique properties

🤔

Challenges

Why 3D data is harder to analyze than 2D images

📊

Analysis

Feature extraction and geometric descriptors

🧠

Deep Learning

PointNet and PointNet++ architectures

❓ WHAT IS A POINT CLOUD?

Definition

A point cloud is an unstructured collection of 3D points in space, each defined by X, Y, Z coordinates.

Data Representation:

Format: 2D array shape (N, 3)
N = number of points
3 = coordinate dimensions (X, Y, Z)

Beyond Geometry:

Intensity: Signal strength from sensors
RGB Color: Per-point color information
Timestamp: When each point was captured
Normal: Surface orientation vector

Acquisition Methods:

LiDAR scanning (airborne/terrestrial)
RGB-D cameras (Kinect, Intel RealSense)
Photogrammetry (SfM)
Structured light

💾 FILE FORMAT COMPARISON

📄

ASCII Formats

.xyz, .txt, .csv

✅ Human readable
✅ Universal compatibility
❌ Large file sizes
❌ No metadata
❌ Slow I/O

📦

LAS/LAZ Formats

.las, .laz

✅ Industry standard
✅ Metadata support
✅ Compression (LAZ)
❌ Not human readable
❌ Limited mesh support

🎮

Binary Formats

.ply, .e57, .obj, .npy

✅ Flexible structure
✅ Mesh + cloud
✅ Fast I/O
❌ Limited standard
❌ Software dependent

🤔 WHY IS 3D ANALYSIS HARDER?

🏗️ UNSTRUCTURED DATA

Points are scattered irregularly in space with no inherent grid structure. Unlike 2D images with pixel grids, there's no built-in neighbor relationship. This breaks traditional CNN architectures designed for regular grids.

🔄 PERMUTATION INVARIANCE

The order of points is irrelevant. A point cloud with points [P1, P2, P3] is identical to [P3, P1, P2]. This property requires special network architectures that can handle permutation-invariant operations.

📏 SCALE AND TRANSLATION VARIANCE

A coffee cup and a building can have identical spatial structure — only coordinates differ. Networks must learn scale-invariant and translation-invariant features to distinguish meaningful patterns from absolute positions.

💾 MEMORY & COMPUTATION

Point clouds can contain millions of points. Processing all points directly requires massive memory and computation. Efficient sampling and hierarchical methods are essential.

🔗 THE NEIGHBORHOOD CONCEPT

Key Insight

A single point's coordinates alone tell you nothing. Feature extraction happens by analyzing its local neighborhood.

📍

Line Pattern

Neighbors aligned in a line → Edge, pole, or thin structure (eigenvalue: λ1 >> λ2, λ3)

📐

Planar Pattern

Neighbors form a plane → Roof, wall, or flat surface (λ1 ≈ λ2 >> λ3)

🌳

Volume Pattern

Neighbors scattered 3D → Vegetation or dense cluster (λ1 ≈ λ2 ≈ λ3)

🎯

Geometric Features

PCA eigenvalues reveal local geometry for classification

💡 Feature extraction = Bridge from raw 3D data to machine intelligence

🌳 SPATIAL QUERIES & KD-TREES

Why spatial queries matter: Efficiently finding neighbors is critical for feature computation. Brute force O(n²) is impractical for million-point clouds.

🎯 K-Nearest Neighbors (KNN)

Goal: Find exactly K closest points to query point
Pros: Fixed neighborhood size, consistent density
Cons: Ignores scale, might include distant outliers
Typical K: 10-50 neighbors
Complexity: O(n log n) with KD-tree

⭕ Radius Search

Goal: Find all points within fixed radius
Pros: Physically meaningful, scale-aware
Cons: Variable neighbor count
Typical radius: 0.1-1.0 meters

Algorithm Insight

KD-tree partitions space recursively, reducing search from O(n) to O(log n) on average. Essential for real-time processing!

📊 COMPUTING GEOMETRIC FEATURES

Method: Principal Component Analysis (PCA)

After finding neighborhoods with KD-tree, extract shape descriptors via eigenvector analysis

Eigenvalue Analysis

Sort eigenvalues: λ₁ ≥ λ₂ ≥ λ₃

Linearity: (λ₁ - λ₂) / λ₁
Planarity: (λ₂ - λ₃) / λ₁
Scattering: λ₃ / λ₁

Classify each point type!

Derived Features

Surface Normal: Smallest eigenvector
Curvature: λ₃ / (λ₁ + λ₂ + λ₃)
Roughness: Deviation from plane
Verticality: Angle with Z-axis

Perfect for semantic labeling!

Typical Workflow:

Point cloud → KD-tree construction
For each point: Find K neighbors or within radius
Covariance matrix computation
PCA decomposition (eigenvalues/vectors)
Feature extraction & classification

☁️ CLOUDCOMPARE: PRACTICAL WORKFLOW

Workflow Steps:

1️⃣ Import Data

Load LAS, PLY, XYZ, or other formats. Preview point count, density, and bounds.

2️⃣ Clean Data

Remove outliers, noise, and erroneous points using statistical filters.

3️⃣ Subsample

Reduce density with spatial subsampling to manage computational load.

4️⃣ Compute Features

Calculate normals, curvature, eigenvalues via PCA on neighbors.

5️⃣ Export Results

Save as CSV, PLY with computed features ready for ML pipelines.

🧠 MACHINE LEARNING APPROACHES

Traditional Pipeline

Manual feature extraction → Classical ML (Random Forest, SVM, etc.) ← Works but requires domain expertise

Modern Deep Learning Pipeline

Raw point cloud → Neural network (PointNet/PointNet++) → Learned features → Classification ← Automatic & powerful

📚 Key Resources for Learning:

📖 Medium Articles

3D Machine Learning course on semantic segmentation and best practices

🔗 OpenPointClass Repository

GitHub project for point cloud semantic classification with multiple networks

📑 Research Papers from Pix4D

CLASSIFICATION OF AERIAL PHOTOGRAMMETRIC 3D POINT CLOUDS

📐 PROPERTIES OF POINT SET IN ℝⁿ

Permutation Invariance

Order of points doesn't matter. {P₁, P₂, P₃} = {P₃, P₁, P₂}. A good network should produce same output regardless of input order.

Translation Invariance

Shifting all points by the same vector shouldn't affect local geometric properties. Networks should be robust to coordinate shifts.

🎯 Global Structure

Aggregate local features via symmetry functions (max pooling, sum pooling) to capture global shape

🔗 Local Dependencies

Nearby points in 3D space relate to each other. Use neighborhoods (KNN, radius) to capture local structure

⚙️ Rotation Sensitivity

Networks are generally NOT rotation invariant. Use data augmentation or alignment preprocessing

📊 Unordered Set

Points are unordered collection, not sequences. Requires symmetric aggregation functions

🔴 POINTNET: REVOLUTIONIZING 3D LEARNING

Key Innovation:

Direct point cloud processing without voxelization

Processes raw (N, 3) point coordinates directly instead of converting to voxels or meshes. Maintains sparse nature of data.

Architecture:

Input: N × 3 point cloud
MLP on each point independently
Max pooling aggregation (permutation invariant!)
Global feature vector
Classification or segmentation head

Advantages:

Simple & elegant architecture
Permutation invariant aggregation
Direct on raw coordinates
Theoretically proven expressiveness

Limitations:

No local structure learning
Sensitive to outliers
Struggles with large point clouds
Limited contextual understanding

💡 Breakthrough: Proved CNNs aren't necessary for 3D learning!

📚 Learn More:

📖 PointNet Implementation Explained Visually

DataScienceUB on Medium - Visual explanation of PointNet implementation

📖 Introduction to PointNet

By itberrios6 on Medium - Comprehensive introduction to PointNet architecture

🟢 POINTNET++: HIERARCHICAL LEARNING

Answer to PointNet's limitations: Learn hierarchical features at multiple scales

Multi-Scale Approach:

Set Abstraction Levels: Recursively partition space
PointNet++ learns: Local features at each level
Coarse-to-fine: Aggregate features bottom-up
Context modeling: Understand neighborhoods + global structure

Key Components:

Sampling layer (fps)
Grouping layer (KNN/radius)
PointNet feature extraction
Feature propagation on upsampling

Advantages Over PointNet:

Learns hierarchical features
Captures multi-scale structure
Better contextual understanding
More robust to noise
Superior segmentation accuracy

Use Cases:

Semantic segmentation (scene labeling)
Instance segmentation
Object detection in 3D
Fine-grained shape analysis

🏆 State-of-the-art for most 3D understanding tasks

📚 KEY RESOURCES

Youtube, Medium

Udemy Course

Youtube

🎉 Thank You!

Questions? Let's explore the fascinating world of 3D point clouds together!